Discriminative relation network for imbalanced graph node classification

Graph-structured data is widely used in biology, chemistry, finance, and other fields. Graph neural networks can learn useful representations from graph-structured data and then various tasks, such as node classification and link prediction. In many graph classification scenarios, the number of samples in each class is usually highly unbalanced, making models to prioritize learning features from the majority class, neglecting minority classes. Traditional solutions to address graph imbalance often involve oversampling, potentially resulting in issues like overfitting and noise introduction. Meanwhile, they fail to leverage the valuable features and structural information in unlabeled data. To solve this problem, this paper designs a framework different from traditional sampling, which first uses an iterative classification algorithm to obtain high-confidence pseudo-labels for unlabeled nodes, and then utilizes a relational network to learn the feature differences between samples, making full use of the information of unlabeled data to effectively generate decision boundaries. Training and learning from three graph imbalance datasets, our experiments demonstrate that the proposed approach outperforms all baseline methods.